How to build AgiBot X1 full-stack open-source humanoid
13 steps to build AgiBot X1 full-stack open-source humanoid, with the source documents.
No price on the official pages; the buy links for the whole-robot and parts packages point to agibotmall.com, which did not resolve on 2026‑11‑03. The open BOM lists parts and quantities but no prices
Buy parts
- Prepare the Logitech F710
The project uses a Logitech F710 controller.
Software
- Open controller mapping settings
Run jstest-gtk to view the controller mapping settings.
Calibrate and test
- Check the default button mapping
View the default controller map; the red characters correspond to controller buttons.
- Reset the simulated robot
In simulation, first switch the robot to ZERO, click reset to make it stand, and then switch to a walking mode.
- Select the real-robot walking mode
Use walk_leg_arm for real-robot walking. Both walk_leg and walk_leg_arm can be used in simulation.
Use it
- Wave with the RB button
In keep, stand, and walk_leg modes, press RB to make the robot wave.
- Use START to enter idle
In any mode, press START to switch directly to idle.
- Use the idle state
idle is the idle state; the joints have no torque output.
- Use the keep state
keep is the holding state; the joints remain at their current positions.
- Use the zero state
zero is the zeroing state; the joints return to zero position.
- Use the stand state
stand is the standing state; the joints return to the standing position.
- Use walk_leg walking
walk_leg is a walking mode with the upper body stationary. Hold LB and push the left stick to walk or the right stick to turn.
- Use walk_leg_arm walking
walk_leg_arm is a walking mode in which shoulder pitch follows the walking motion. Hold LB and push the left stick to walk or the right stick to turn.
Common problems
Parts
- PowerFlow joints (development-guide counts): R86-2 ×9 (48V, 80 N·m peak), R86-3 ×6 (48V, 200 N·m peak) and R52 ×10 (48V, 19 N·m peak), all with hollow cabling, built-in FOC drives and multi-turn absolute encoders
- L28 linear actuators ×4 (24V, FDCAN, 110 N of thrust); OmniPicker adaptive grippers ×2 (0.4 kg, 30 N of grip)
- Xyber-DCU v18 domain controllers ×2 (EtherCAT to 3 FDCAN buses), a MiniPC industrial computer, YIS320 attitude sensors ×2 and a D435 camera
- A lithium battery with a 13S 54.6V 5A charger (toolkit list), two soft-start boards, 19V and 24V buck modules, a rocker main switch, a remote switch and an XT60 connector
- Structure: machined TC4 titanium, 7075 and 6061 aluminium and 0Cr17Ni4Cu4Nb stainless parts, ABS shells and rubber soles; crossed-roller, four-point-contact, rod-end and spherical bearings
Tools
- Machining and shell production (SolidWorks 2022 sources, STEP and PDF drawings)
- From the toolkit list: a robot hoist rated for 300 jin (150 kg), a 1-tonne 1.5 m lifting sling, four M4 eye bolts with snap hooks, plus zeroing blocks and a neck handle
- 0.2–1.2 N·m and 1.0–6.0 N·m torque wrenches, hex keys and bits, Loctite 243 threadlocker and an electric screwdriver
- A soldering station, heat gun, wire strippers, an Ethernet crimper and heat-shrink; a multimeter to check the battery for shorts before power-on
- A Windows PC for REF-CLI; an x86 PC with real-time Ubuntu 22.04 for the bench test
Skills
- Setting actuator CAN IDs, zero offsets and firmware versions in the REF-CLI command line, and zeroing the whole robot
- Building and deploying the AimRT software with a Linux real-time kernel, GCC 13, CMake 3.26+, ONNX Runtime and ROS 2 Humble
- Reinforcement learning in Isaac Gym Preview 4, sim2sim checks in MuJoCo and JIT/ONNX export
Software
- agibot_x1_infer (MulanPSL-2.0): RL inference, DCU joint driver and gamepad modules on AimRT, with ROS 2 simulation and examples
- agibot_x1_train (BSD-3-Clause): locomotion training built on legged_gym, rsl_rl and humanoid-gym, sim2sim to MuJoCo and JIT/ONNX export
- AimRT (MulanPSL-2.0): AgiBot's open-source robot middleware
- The REF-CLI actuator tool and the xyber_dcu_test bench-test software (downloads in the development guide)